Open models debate: diverse specialization, not today's scores, decides open source's value
willcb · x · 2026-09-15
- Replying to @tenobrus, who doubts open models can make a real positive difference, willcb argues believing in open source requires just two things: training matters, and diverse specialization is economically useful — with all of modern economic history backing the latter.
- The key difference between open and closed static models is not current access but the ability to keep training. In the limit, flops/data/knowledge/skill/speed/cost/volume all trade off; "capabilities" is a point-in-time phenomenon, and the evolved transformer will saturate and specialize like the evolved human brain.
- Follow-up: he finds it wildly inefficient that in 20 years most inference flops might run on a single static weight set with knowledge retrieved in context; a many-model equilibrium is just like the human world — powerful players have more influence, but that's the status quo.
Related event: Researchers Clash Over Whether Open-Weight Models Help or Harm(5 posts)→
More from AGI Musings
- Stuart Russell: AI safety needs concrete goals, not just slower timelines — nordicinst · 2026-09-15
- A decade of doomsday warnings failed to slow the AI race, Guardian analysis finds — nordicinst · 2026-09-15
- Dario Amodei's answer on whether AI could kill us all by 2030 read as 'yes' — zetalyrae · 2026-09-15
- Scaling plateau is 3-4 OOMs away, and at 10x/year superintelligence lands by 2030 — willcb · 2026-09-15
- davidmanheim: we can't yet steer AI well, frontier labs bet on fixing it later — davidmanheim · 2026-09-15
- Alignment researcher: broadly adopting weakly aligned strong AI would be disastrous — davidmanheim · 2026-09-15